Meta backs five U.S. universities studying AI in medicine and science

Meta is putting its AI models into university-led studies of scientific reasoning and clinical care. On October 28, 2026, the company announced support for five U.S. universities, combining model access, researcher collaboration and funding for select projects. The studies will examine questions ranging from predicting unfamiliar chemical reactions to helping doctors interpret patient records.

The participating institutions are Caltech, Ohio State University’s College of Medicine, Louisiana State University, Indiana University and UTHealth Houston. Their work spans medicine, mathematics, physics, chemistry, biology and computer science, with different forms of support for each institution and project.

Researchers will use Muse language models alongside image-analysis tools such as Segment Anything Model and DINO. Caltech researchers will access Meta’s frontier Muse models through its first-party API, a software interface for accessing the models.

Can the models produce science that holds up?

Caltech’s projects target a difficult distinction: whether a model can contribute to original research rather than reproduce familiar patterns. Mathematician Tony Yue Yu is building MathDB Frontier, a benchmark based on a database of more than 86,900 research-level open problems. It will test work such as forming conjectures, finding supporting mathematical results and generating machine-checkable proofs.

• In chemistry, Sarah E. Reisman is using unpublished electronic lab notebooks to build a reaction-prediction benchmark. The aim is to distinguish predictions on genuinely new data from recall of published results.
• In physics, Nai-Chang Yeh is developing a framework that makes models construct candidate descriptions of physical systems under explicit constraints, testing whether they can move beyond plausible but ungrounded descriptions.
• In computer science, Anima Anandkumar is developing TorchLean, a framework using the Lean 4 proof system. It will let AI agents construct, test and repair scientific and machine-learning outputs whose correctness can be formally checked.

Testing medical assistance—and doctors’ reliance on it

At Ohio State, one team will combine routine abdominal CT scans with medical records collected over time and ECG data when available. The goal is to predict future risk of heart failure with preserved ejection fraction before clinical diagnosis. The approach draws on information already collected during routine care rather than relying only on a dedicated heart evaluation.

Another Ohio State project will use Muse models to draft chest X-ray reports using multiple kinds of input and comparisons with prior exams. Researchers will also adjust AI-generated accuracy scores using a fixed set of reports annotated by people, addressing bias when one language model evaluates another’s work.

A related study will examine how radiologists use those drafts in a simulated clinical setting. It will distinguish rejecting correct AI content from accepting incorrect content, then connect those behaviors to the kinds of errors the model makes. The study therefore examines the human response to an AI report, not just the report’s accuracy.

Ohio State will also develop a clinician-guided system of AI agents for Alzheimer’s diagnosis. It aims to reconstruct cognitive history, weigh clinical evidence, identify uncertainty and recommend informative next steps, while keeping clinicians in charge of decisions.

At LSU, the Eyes-On project is developing camera-based monitoring for displaced tracheostomy tubes and signs of distress in infants who cannot reliably vocalize. It is exploring Segment Anything models for tracking and is designed for bedside cameras and local computing. A proof of concept has been published; further work remains.

Faculty set the questions

Meta says faculty brought their research questions to the company, designed the studies and chose how to use the tools. Meta helped identify suitable models and technologies. The support package does not mean every project receives funding: financial backing is reserved for select projects, and Meta did not disclose an amount.